BETA
Instruction tuning fine-tunes a base model on instruction-response examples so it follows user directions instead of merely continuing text.
A base model completes text; ask it a question and it may respond with more questions, because that continuation is statistically plausible. Instruction tuning trains on large datasets of prompts paired with high-quality responses, teaching the model that input is a request and output should be a fulfillment.
It is the first stage of post-training, usually followed by preference optimization such as RLHF, and together they turn a raw checkpoint into the instruct or chat variants that providers actually serve. Open-weights releases typically publish both base and instruct versions, with base models reserved for teams doing their own post-training.
Every commercial API model is instruction-tuned, so buyers rarely face the choice directly — but the quality of a model's instruction following determines how elaborate prompts must be, and shorter effective prompts mean fewer billed input tokens.
Last revised 2026-07-05 · All glossary terms → · Live AI model pricing →